Qwen2.5-Gutenberg-Doppel-14B Overview
This model is a 14.8 billion parameter language model developed by nbeerbower, based on the Qwen2.5-14B-Instruct architecture. It has been fine-tuned using the ORPO (Optimized Reward Policy Optimization) method over three epochs, leveraging two distinct Gutenberg-derived datasets: jondurbin/gutenberg-dpo-v0.1 and nbeerbower/gutenberg2-dpo. This specialized training aims to enhance its text generation capabilities and instruction following.
Key Capabilities & Performance
The model demonstrates notable performance across several benchmarks, as evaluated on the Open LLM Leaderboard:
- IFEval (0-Shot): Achieved a strict accuracy of 80.91%, indicating strong instruction-following ability.
- BBH (3-Shot): Scored 48.24% normalized accuracy on Big-Bench Hard, reflecting its general reasoning skills.
- MMLU-PRO (5-shot): Recorded 43.57% accuracy, showcasing its performance on advanced multi-task language understanding.
It supports a wide array of languages, including English, Chinese, French, Spanish, German, and more, making it suitable for multilingual applications.
Good For
- Instruction Following: Excels in tasks requiring precise adherence to given instructions.
- General Text Generation: Suitable for generating coherent and contextually relevant text.
- Multilingual Applications: Its broad language support makes it versatile for global use cases.
- Research and Development: Provides a strong base for further fine-tuning or experimentation, particularly for tasks benefiting from its Gutenberg-derived training.